Integrated photonic neural network with on-chip backpropagation training.

Area of Science:

• Photonics
• Artificial Intelligence
• Deep Learning

Background:

• Scalable integrated photonic neural networks (PNNs) require high-quality training for robust performance.
• Current PNN training often relies on digital computers or less versatile gradient-free methods due to the lack of on-chip activation gradients.
• Device variations and environmental factors impact the performance of digital-computer-trained PNNs.

Purpose of the Study:

• To demonstrate an integrated photonic deep neural network trained end-to-end using on-chip gradient-descent backpropagation.
• To enable all-optical linear and nonlinear computations on a single photonic chip for scalable and robust training.
• To achieve PNN training that matches digital model accuracy and robustness without external digital computation.

Main Methods:

• Development of an integrated photonic deep neural network capable of performing all computations on-chip.
• Implementation of end-to-end training using on-chip gradient-descent backpropagation.
• Testing the photonic neural network on two nonlinear data classification tasks.

Main Results:

• Successful demonstration of an integrated photonic deep neural network trained with on-chip gradient descent.
• Achieved scalable and robust training of PNNs despite fabrication-induced device variations.
• Matched the accuracy (over 90%) and robustness of a reference digital model in classification tasks without digital computer assistance.

Conclusions:

• Integrating backpropagation training directly onto photonic chips offers a scalable and robust solution for photonic computing.
• This on-chip training method overcomes limitations of current PNN training approaches.
• Enables generalization to various PNN architectures, paving the way for future scalable and reliable photonic computing systems.